Half of every team has not started

Four economists published a task-level map of AI use at work in August, and one number in it surprised me. Across the tasks people actually do, only 2.8% show more than half of workers using AI, and not one task reaches 70%. The technology touches about four occupations in five, and inside them most of the work has not moved. I have been quoting adoption percentages for two years and I think I was reading them wrong.

A few of the things I read last week:

• Bick, Blandin, Deming and Schumacher built the first task-level adoption indexes and found use is broad and thin: genAI reaches around 80% of occupations, fewer than half of workers use it in most of them, and only 15% of occupations get past 70%. (NBER)
• Salesforce reported Agentforce revenue above $1.5 billion and 3.2 billion "agentic work units" in one quarter, while about 96% of its earnings beat came from paper gains on its Anthropic stake rather than from selling software. (Memeburn)
• Procapita puts Gulf enterprise adoption at 84%, up from 62% two years ago, and reports one global organisation restructuring 16,000 roles to fit AI-supported work. (Consultancy-me)
• Across 2.4 million workers and 16,753 skills, Carpanelli, Duszynski and Stephany find depth sorts people into better-paid occupations while breadth moves them into roles with less automation risk. (arXiv)
• Simon Willison says the working skill with coding agents is instructing them confidently and then verifying confidently that they did what you asked. (simonwillison.net)

None of that is an adoption problem. It is a question about what happens inside a role once the licence has arrived.

The percentage counts who has an account

The adoption figure is a good number for deciding whether to buy something and a poor one for deciding what to do next. Bick and his colleagues cut adoption by task instead of by person, and the picture changes shape. AI turns up somewhere in four out of five occupations, which is the headline everyone repeats. Inside them, most people doing most of the work have not started, and the occupations past 70% are mostly in computing. They also found that vendor estimates run high, because a chat log labelled "edit documents" describes work only 2.4% of workers actually do, against the roughly 15% the vendors infer.

What stays with me is the shape of the distribution rather than its size. If one team holds people who have rebuilt their week around the model and people who have opened it twice, the average describes neither of them, and the average is what goes in the board pack. As we wrote on Wednesday, the ECB has half of Europe's workers using AI and a median saving of three hours a week. That number has the same problem, and I used it anyway.

Conversion is a decision somebody has to make

Salesforce counted 3.2 billion agentic work units last quarter. That is a real number describing real activity, and it says nothing about whether the activity was worth doing, which is roughly what Sangeet Paul Choudary meant on 23 August when he wrote that strategy should follow the constraint rather than celebrate the newly abundant capability.

The organisations that seem to have got past this are not running deeper training. Procapita's Gulf figures describe one organisation moving 16,000 roles to match how the work now happens, which their analyst calls a deliberate rebuild rather than downsizing. I would want to see that up close before calling it a model, but it is at least the right unit of change. Willison's point about agents is the same idea at the desk: instructing well and verifying well are things a person learns beside somebody who already can, which is a coaching structure rather than a course.

The fair objection is that there may be nothing to convert. An NBER survey of about 6,000 executives in four countries, published in February, found nearly 90% saying AI had made no difference to productivity or employment in three years, and the average executive who uses it spends about ninety minutes a week doing so. Read that way, shallow use is people sensibly declining a tool on work it does not help with, and pushing for depth would only produce more theatre. I take that seriously; I would rather a team used AI on four tasks properly than on forty badly. What it does not explain is the variation Bick and his colleagues found between workers doing the same job. If the task were the whole story, those people would look alike.

Ingka Group retrained 8,500 call-centre staff into remote interior design advisers when its chatbot took 47% of customer contacts, and the new service brought in €1.3 billion; this spring the group and its franchisor still cut around 1,650 office roles. The redesign was real. It was also a decision taken once, not a standing commitment.

If you work in L&D, HR, or transformation

We are usually handed the adoption number and asked to raise it, and I understand why that is the request. I would change the unit instead. Take one team and one task they do every week, and find out how many of them use AI on that task rather than on anything at all; the gap between that answer and the headline figure is the size of the real job. Then put two people from the same role who are far apart on it in a room together, and ask the one further along to show the other how they check the model's work. It costs an hour and it is closer to capability than another rollout.

The provocation

Pick one task your organisation depends on every week. Do you know how many of the people who do it are using AI on that task, and if not, who in your organisation should be finding out this week?